As reported on November 18, 2026, Shopify had rebuilt its internal Overexposed scanner to check whether Kubernetes Ingresses on GKE and Cloud Run services could be reached from the internet without authentication. The earlier version relied on infrastructure code; the rebuilt system combines Google Cloud Asset Inventory data with anonymous HTTP requests.
Shopify said the code-based scanner had three blind spots. Directly configured proxy-authentication annotations could produce false positives. Admission controllers could change settings before a cluster accepted them—and not every cluster ran the same controllers. Ingresses created by other repositories or platform tools could be missed entirely. The lesson, according to the account of the project, was that a repository shows intended configuration, while the cluster shows what was accepted.
The current scanner runs on a schedule. It queries Cloud Asset Inventory for accepted GKE Ingresses, Cloud Run services and load balancers, and separately obtains access-policy information without cluster credentials. That inventory also revealed Cloud Run services in projects the first scanner had not covered.
For each candidate, Overexposed sends an anonymous HEAD request, falls back to GET if needed, follows up to three redirects and checks every step in the redirect chain. It also tests path variants such as /ADMIN, since path rules can be case-sensitive. Results are classified as exposed, expected, protected, not found, unreachable or indeterminate. “Expected” means a public endpoint exactly matches a reviewed allowlist entry in code, including its hostname and path. Those endpoints are still probed on every run; entries are flagged as stale if the endpoint becomes protected or disappears.
The scanner does not treat an HTTP status code as a verdict by itself. A 404 behind an Ingress can mean the proxy responded but the requested path did not exist. For Cloud Run, Google’s frontend can return 404 before the application runs, so Overexposed may classify the result as indeterminate. A 403 is not automatically proof of protection either. For Cloud Run, the scanner checks ingress settings and IAM: a service is public only if allUsers has the invoker role or the invoker check is disabled. For services behind load balancers, URL Maps identify the backend for each path and locations requiring authentication mechanisms such as Identity-Aware Proxy.
In one July scan, roughly three-quarters of results were indeterminate. Shopify grouped those results by cause, added definite classifications where evidence supported them and replayed saved scan results to test rule changes before release. After several weeks of tuning, the indeterminate share of Ingress results fell by about 80%. Where evidence remains insufficient, the scanner retains an indeterminate result rather than assuming a service is protected. It records the verdict, its reason and the responsible team, but not response bodies or credentials.
Because the scanner makes outbound requests to arbitrary hostnames, Shopify also hardened it against server-side request forgery. It checks every IP address returned by DNS, rejects private and loopback addresses, connects directly to a checked IP to guard against DNS rebinding, and repeats those checks on redirects. The project’s authors also recommend restricting the scanner’s outbound network access, including access to cloud metadata endpoints.
Shopify built Overexposed using AI coding agents and AI code review. It uses mutation testing on security safeguards to catch tests that pass while checking the wrong behavior, and keeps a canary target for each verdict category. The approach illustrates the distinct roles of configuration checks, cloud inventory and active probing: inventory can find running resources that repository scans miss, while an anonymous request can provide evidence about what an outside visitor can reach. Neither a configuration flag nor an ambiguous HTTP response should be presented as conclusive proof of protection.